Dwarkesh Patel Podcast - Grant Sanderson – AI and the future of math
What this covers
Grant Sanderson and Dwarkesh Patel use mathematics as a lens on where AI is strong and where it stalls. They start with spikiness: AI reached IMO gold largely because a brute-force solver dispatches geometry in seconds, while combinatorics holds out, and the unevenness repeats inside math's own sub-areas. They then argue over whether Lean formalization is the real driver. Patel calls it overrated, pointing to a natural-language counterexample on the Unit distance conjecture and to DeepSeek's verifier checked by a meta-verifier. Sanderson counters that Lean could let a system run for years, AlphaGo-style, extending Mathlib unsupervised.
The historical middle section explains why some value resists quick feedback. Galois was rejected by the academy, and about twenty years passed before Liouville recognized his work and twenty more before Jordan's modern treatment. Group theory took roughly a century to prove its worth, through cryptography and Gell-Mann's prediction of quarks. Sanderson contrasts building whole theories, as Fermat's Last Theorem required with elliptic curves and modular forms, against sudden cross-field links like Hugh Montgomery and Freeman Dyson finding random-matrix statistics in the zeta zeros. He also discusses the Langlands program and an AI solution to Erdős problem 1196.
The contested turns come last. Sanderson wonders whether next-token autoregression suppresses the improbable connections that matter most. Patel suggests that many breakthroughs are connections between existing fields. They link models' weak theory of mind to poor spaced-repetition cards, and tie sample efficiency to AI's reliance on replayable environments. Sanderson also separates unsolved problems from unsolved explanations, such as forcing and the continuum hypothesis.
Sanderson argues that math is AI's most advanced test case because its verifiability and grindability make progress legible, but the hardest mathematical breakthroughs (new conceptual frameworks, good definitions, and explanations) resist benchmarking and RL training, so mathematicians' roles will shift toward curation and explanation rather than disappear.
- Math progress is spiky: contest problems like IMO geometry fall to brute force while conceptual breakthroughs like Galois theory take a century to be verified
- Verifiable, grindable outcomes make RL work, and natural-language verification with meta-verifiers can substitute for Lean
- Human curation, social motivation, and trust remain distinct from raw explanatory or proof ability
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Group theory's usefulness took roughly a century to verify, passing through cryptography and physics, so the value of Galois's insight cannot be checked by a short verification loop and resists quick RL-style rewards.
“potentially that verification loop is a hundred years long.”
Much of modern mathematics is organized around the Langlands program, a research ethos of preemptively finding connections between fields that often knocks down specific problems, unlike work that targets one problem at a time.
“It’s not even a field of math so much as it is a research ethos.”
Galois's manuscripts were rejected by the academy and his brother and friend's efforts failed; it took about twenty years for Liouville to recognize them and about twenty more for Jordan to produce a modern treatment, showing how long a valid idea can wait for recognition.
“it was another twenty years until Liouville sees these notes”
Murray Gell-Mann used a purely group-theoretic argument about how symmetry groups break down to anticipate the existence of quarks, an application of group theory that came long after its founding.
“He anticipates quarks based on a purely group-theoretic question.”
Beyond unsolved research problems there are unsolved expository problems: results proven as true but not understood, such as the independence of the continuum hypothesis via forcing, and the gap between proof and explanation is the key distinction.
“I want to propose the idea of an unsolved expository problem. Sure, we’ve proven it, but we don’t really know why it’s true.”
Lagrange introduced the instinct that symmetries of polynomial roots were the right object of study and that quintic solvability might be impossible, but he solved nothing, so the value of his question was only recognized later.
“there wasn’t any result that Lagrange came to. He didn’t solve the problem, so it wasn’t a case of knowing it was the right question to ask based on a solution.”
AI reached IMO gold-level performance largely by solving geometry problems with a brute-force solver in seconds, while combinatorics remains the holdout; in 2024 the gold would have come sooner had the test contained more geometry problems, since the category split was a toss-up.
“Geometry, it just solves it in nineteen seconds since 2024 because it’s a brute force solver.”
There is no single correct heuristic for doing science, so progress depends on multiple independent research programs with different heuristics, though Einstein's bias that God does not play dice shows a strong bias can halt progress in other areas.
“Which goes to show you that there’s not a correct heuristic for science. You just need multiple independent research programs with their own heuristics.”
Hugh Montgomery found that the pair correlation of Riemann zeta zeros matched a formula that Freeman Dyson recognized from random Hermitian matrix eigenvalue statistics, linking number theory to physics and suggesting a cross-field bridge.
“Freeman Dyson, a physicist, is like, 'I know that expression. That expression comes up in studying the eigenvalues for random Hermitian matrices'”
Proving Fermat's Last Theorem required first building two large theoretical mountains, elliptic curves and modular forms, before the right question could connect them, illustrating how deep proofs can need whole new theories.
“Both of those mountains have to be built before you can ask the right question that connects them.”
Autoregressive next-token prediction is a strange way to generate mathematics, because the substantive cross-field connections are by nature unlikely next tokens, so training on predictable continuation may suppress them.
“But the connection where all the substance is going to come from is, by its nature, a very unlikely one.”
Dwarkesh finds LLM explanations of human writing better than the original, yet LLMs struggle with the mentalizing needed to write good spaced-repetition cards, since writing for a reader requires projecting that reader's mind months later, which models do poorly.
“The key constraint, it seemed to me, was that writing a good card is about projecting somebody’s mind in three months.”
Deep learning needs many parallel rollouts to acquire a skill because sample efficiency has not been solved, which is why the sparse supervision transmitted through a straw limits training on real-world tasks.
“The reason you currently need to do so many parallel rollouts to learn a skill with deep learning is that we haven’t solved sample efficiency.”
Current deep learning needs huge numbers of parallel rollouts because sample efficiency is unsolved, and this forces AI to rely on containerized, replayable environments such as code and math rather than the messy real world.
“The reason you currently need to do so many parallel rollouts to learn a skill with deep learning is that we haven't solved sample efficiency.”
Humans understand others' emotions partly by mimicking facial expressions with their own muscles, and since models lack face muscles and have a different brain architecture, their poor theory of mind is unsurprising, like an alien trying to empathize.
“The thought is that part of understanding the emotion you’re looking at is doing it yourself. At a facial level, you’re moving your face muscles.”
The AI solution to Erdős problem 1196 on primitive sets brought in a Markov chain approach and the von Mangoldt function from another field, a small cross-field idea that humans can readily parse because only its start and endpoints need explaining.
“Those are going to be very human-parsable, because all you have to do is show the start and end point of what those connections are.”
Lean formalization is probably overrated as the driver of AI math progress, since the unit distance counterexample was produced in natural language without Lean, and process-level correctness matters less than a grindable, verifiable outcome.
“It didn’t have any Lean in it. I think the process-based supervision that Lean provides... seems less relevant than just having this grindable outcome that is verifiable.”
DeepSeek's math work shows that natural-language process verification, with a verifier trained and checked by a meta-verifier, can provide good feedback, so formal systems like Lean are not required for process-based supervision.
“the verifier is trained by a meta-verifier that makes sure that for all the problems they’re training this model to solve... the verifier is giving good feedback. It works.”
Patel suggests that in the limit many of the biggest breakthroughs, such as general relativity connecting Riemannian geometry and special relativity, are essentially connections between existing fields, so better AI connectors could produce major breakthroughs.
“With general relativity, you’re just connecting together Riemannian geometry and special relativity.”
Lean could let an AI run continuously and extend Mathlib with no human check-ins, similarly to AlphaGo exploring Go, because Lean provides an automated verifiable reward for proofs, so compute can be poured in for years and the result reviewed later.
“you could imagine having a basically endlessly running program that’s constantly trying to extend Mathlib.”
The time needed to recognize a valuable idea depends on its type: relativity was accepted quickly, while group theory and similar conceptual frameworks required a verification loop of about a century.
“With Einstein and GR, people could feel this was a good theory right away.”
AI's spiky capability profile recurs at finer scales within mathematics itself, so some sub-areas of math are much easier for AI than others, meaning the spikiness is fractal rather than a single frontier.
“But there’s a fractal nature to that spikiness, because when you zoom into the specific progress within math, you have some things that are a lot easier than others.”
Theorem-proving receives most credit but parasitizes definition-making; if AI automates proofs while humans supply insight, credit allocation breaks, as Bessi's 'Fall of the Theorem Economy' predicts.
“The theorem-proving stuff is what gets all the credit, but it’s really a parasite on the coming-up-with-the-definition stuff.”
The next meaningful AI milestone in math is generating interesting problems and new conceptual objects that unify fields, not just solving posed problems, since motivating why a problem matters is what drives breakthroughs.
“coming up with interesting problems in the first place, and the other is coming up with new kinds of objects or conceptualizations that create or unify fields.”
Failed AI reasoning often comes from a bad chain of thought that is hard to escape, so restarting with fresh or deliberately different context, as humans do when they step back to try the opposite, is a useful systematic technique.
“Sometimes when an AI is failing, it’s because it gets into a bad chain of thought and it’s really hard to get it out. So you say, 'I’ll just start again.'”
Once AI can prove and explain results, the mathematician's role shifts toward curation, deciding which ideas are worth pursuing and presenting, much like an art museum curator, because people trust a human's choice of what is interesting.
“So my role, and arguably that of other mathematicians, might actually just shift subtly into that curation direction of what ideas are worth pursuing.”
The economic value of pure math is spiky: algebraic number theory unlikely unlocks economic gains, while PDE and dynamics insights can produce incremental engineering improvements such as those that saved Boeing billions via simulation, and some leakage into engineering is likely.
“With progress in algebraic number theory, it feels unlikely that that then unlocks something.”
Anything that can be turned into a benchmark can also be a training environment, so tasks like conjecture generation that resist benchmarking also resist current RLVR training, which requires a fundamentally different approach.
“There’s really no fundamental difference between a benchmark and a training environment.”
AI minds can be parallelized, scaled, cloned, and merged in ways humans cannot, so a system applying its capability across all accessible problems can overcome the lack of a single genius, and massive spending makes quantity have a quality of its own.
“Whatever level of capability they have, it’s not just one idiosyncratic genius in the history of mathematics who makes a few connections and then dies in a duel.”
Conjecture-generating ability cannot be cleanly benchmarked the way a solved problem can, so it will show up as a subjective tone shift in how mathematicians describe AI being useful for choosing research directions.
“I think the way you’d measure conjecture-generating ability is going to be more subjective, based on that tone shift.”
A Riemann solution could come as a lightning bolt connecting two fields, a new mountain of theory, or brute-force hustle over a very long chain of reasoning; these differ in how easily humans can digest the result, with brute force raising the worst digestion problem.
“if the nature of the progress was just raw hustle—a super long chain of reasoning with no new theories—then you would have that worry of this whole digestion process.”
The attempted proof of the abc conjecture via inter-universal geometry was theory-building rather than a long chain of reasoning, so it took years of effort to parse, and the biggest danger of AI is producing a similarly plausible-looking but wrong new mountain that consumes years of human checking.
“The biggest fear would be that an AI does that, and then much like the abc conjecture, people work for years to go up the mountain”
Credit assignment is solvable in code because deterministic rollouts from the same start point let the diff explain success, but it becomes much harder when tasks start from different points and the world changes.
“If you have situations that are starting off at different starting points, this credit assignment problem becomes much harder to solve.”
Formal verification gives mathematicians a green checkmark that a proof is correct, sparing them from reviewing AI-generated papers where even a 99 percent accuracy rate leaves each paper costly to vet, so formalized outputs would be hugely valuable.
“Even if 99 out of 100 are right, I don’t know if it’s worth my time because it’s really labor-intensive to find what that error would be.”
LLM verifiers can discriminate a B essay from an A essay but fail to tell an A essay from one that is actually insightful and readable, and they tend to prefer uninsightful writing, so reward hacking derails writing progress.
“they’re not actually good at discriminating between an A essay and a thing you actually want to read... They actually end up preferring uninsightful pieces of writing.”
AI systems trained similarly tend to think the same way, producing homogenized outputs, and this entropy collapse can be countered by systematically trying both a statement and its negation or assigning different biases to different agents.
“a common concern people have about AIs is this entropy collapse where they all think the same way, because they’re trained in similar ways.”
In mathematics, a new idea frequently fails ordinary human review and survives only because one person recognizes its value, so the verifier is often a single insightful reader rather than a consensus.
“A lot of the time, it’s not even passing the verified reward of human reviewers. It gets on someone’s desk and they say, “I don’t really know if there’s anything here.” You have to have this one person recognize it.”
AI progress in math is fast not only because outcomes are verifiable but because the domain is grindable: deterministic containers and diffs let RL run many parallel rollouts and solve credit assignment, which computer use and the real world lack.
“It’s not just verifiability; it has to be grindable.”
Making a small connection between two fields (a lightning bolt) is different from building a new theory (a mountain): lightning bolts are easy for humans to verify and digest, while new mountains require much more effort to understand.
“If the character of it is mountain building, you have to put in a lot more time to understand that new mountain that was built, because it's a new thread, not just a lightning bolt between them.”
Even if automated engineers can make predictive advances without understanding, mathematics' end goal is human understanding, so there will always be a need for the concision function that distills complicated reasoning into a simple, understandable form.
“presumably, the end goal is understanding, human understanding.”
For getting AI to make cross-field connections, reasoning about the training data and environments is more productive than reasoning about architecture or loss function, because the environment determines what behavior gets incentivized.
“I think it’s more productive to reason, instead of architecture or even loss function, about data.”
Autoregressive next-token generation is a strange way to produce insight, because the valuable connections are by nature unlikely next tokens, and standard RL does not specifically upweight such low-probability leaps, so the system may hold intelligence locked inside an awkward interface.
“the connection where all the substance is going to come from is, by its nature, a very unlikely one.”
To reward the Galois-like instinct for a valuable concept rather than just solved problems, one would need a verifiable proxy for conceptual smallness, such as Kolmogorov complexity, since compact, predictive ideas feel more intelligent.
“I wonder the extent to which you can give some kind of verifiable reward around not just whether you solved it... but around the smallness of the concepts required to do it.”
What makes writing valuable is generating new insight through deliberate, unpredictable choices, not distilling existing ideas, so an LLM summary of a book is not a substitute for the author's exploration and motivated narrative.
“It’s about what the insight is.”
Writing progresses more slowly than code or math because in writing the text is the end product, so every sentence matters and it cannot be sloppy in the way code can be if it still functions, while code and lemmas are modular intermediate steps.
“Whereas in writing, the end product is directly the thing the AI is producing. Each paragraph, sentence, and word matters because that is the substance.”
Automated provers could exhaustively search possible axiom systems, most yielding trash but occasional rich islands, and such islands could later be retroactively motivated, as with group axioms that look arbitrary until their link to symmetry is understood.
“every now and then, there would be this little island of a completely different type of axiom system”
LLMs are sycophantic and run off with the student's question rather than reframing it, whereas the best human teachers jujitsu a student's creative misconception into the correct structure, so LLMs fall short of top explainers.
“They aren’t reframing your question. Instead, they kind of run off.”
Single-author expositions such as the Stanford Encyclopedia of Philosophy or the Princeton Companion to Mathematics are better for learning than Wikipedia, because one author can craft motivation and deliberately leave things a little wrong to correct, whereas crowdsourcing converges to a local minimum where every sentence must be correct.
“The difference there is the articles are deliberately written by one individual who tries to actually craft a motivation around it.”
If AI accelerates pure math without affecting applied fields, mathematicians may have to confront that grant proposals promising applications such as elliptic-curve progress for cryptography were overstated, showing math may be divorced from physical applicability.
“Every time we wrote those grant proposals and said, 'Trust us, the elliptic curve progress is going to help with cryptography,' it shines a light on the fact that maybe it doesn’t.”
Grant has revised his view: he used to expect AI would be automated theorem provers while mathematicians shifted to explaining results, but he now suspects AI will also be better than most humans at explaining and distilling, so explanation will not remain a human niche.
“Now I suspect that actually they’ll also be quite good at doing that, probably better than most humans are at explaining and distilling.”
Students choosing mathematics careers think little about where funding and salary come from and what value they add, relying on the pleasure of engaging with math rather than analyzing the institutional logic of brand value, grants, and teaching.
“you should understand where the money is coming from, what value you’re actually adding, and the connection between those two. A surprisingly small amount of thought is put towards that, especially by students.”
AI has been slow at computer use because websites use bot detectors that block parallel rollouts, and because deep learning lacks sample efficiency, so the many rollouts needed to learn a skill cannot be run on real sites.
“it’s very hard to run a thousand parallel rollouts of the same checkout flow on Amazon. You’ll get shut down”
If AI solved the Riemann hypothesis by connecting two expert fields with a single lightning-bolt insight, that skill would be quite distinct from what white-collar work requires; but if it required building an entirely new theoretical mountain, that level of creative theory-building would likely spread into the broader economy.
“If the solution to the Riemann hypothesis involved building a new mountain, that’s a kind of skill... that’s just such a level of intelligence that it would be surprising if it didn’t permeate into other aspects of the economy”
Good writing requires deliberately choosing unpredictable, novel moves that contradict typical production, and this is hard for autoregressive generation because it has to know exactly when to depart from the expected next token.
“It’s knowing exactly the correct point when you want to make an unpredictable move, and that that’s going to be what’s more insightful.”
Even a thousand-page proof leaves the goal unmet, because the end goal of mathematics is human understanding; prediction alone, as with engineers building rockets without understanding, does not satisfy that goal.
“presumably, the end goal is understanding, human understanding. Even if you do have some thousand-page proof of some math thing or some grand new physical theory, the goal is understanding.”
In choosing courses or books, students should weigh the educator or author's quality and resonance more than their prior topic interest, since interests are arbitrary early on, and follow authors rather than subjects.
“care a little bit less about your preexisting interests, because they’re kind of arbitrary right now, and care a little bit more about whether the person teaching it is a good educator”
The most productive LLM use for learning is to ask which human-written resources to read, using the model as a souped-up search that points to the right book, article, or video, because LLM explanations are amazing but can be confidently wrong, as when a fabricated video attribution was given.
“So often I like to just ask an LLM, 'Who should I read?'”
The most productive learning combines a human-produced artifact that organizes concepts in the right order and motivates each next idea with an LLM used only to prune the branches the book has identified, since LLMs cannot independently structure a field.
“Then you use the LLMs to just do a little bit of pruning around this branch that the book has identified.”
Mathematicians could design partially synthetic problems that require connecting two fields and remove assumptions so the AI must reach the answer, creating an RL environment that incentivizes cross-field connection regardless of the loss function.
“It's really about, can you come up with an environment that incentivizes this ability?”
Teaching is one of the most stable jobs in a post-AGI world because it is relational and coaching-based, and affluent parents will spend abundance on good educators, so students should consider math education as a durable career path.
“I actually think teaching is one of the most stable post-AGI jobs that there is, because it’s so relational.”
If solving the Riemann hypothesis required finding a cross-field connection, that skill is distinct from what white-collar work demands, because LLMs are expert in many fields and should be able to see such similarities without a lucky lunch conversation.
“They’re experts at quantum physics. They’re experts at analytic number theory. They should be able to see that similarity in a way that doesn’t require Montgomery and Dyson to be having lunch”
People who produce genuinely novel insights, such as Einstein, Shannon, and Feynman, are also typically very clear expositors, suggesting that the same cognitive capacity drives both new thinking and good explanation.
“There seems to be a really strong correlation between the people who come up with genuinely novel insights and the people who are actually quite clear in their communication of it.”
LLM explanations are amazing but resemble Wikipedia in being a local minimum where every sentence must be correct, so they lack the deliberately crafted wrong-then-corrected motivation of a single-author exposition.
“LLM explanations feel to me at the moment a lot like Wikipedia, which is to say, amazing.”
If AI produces a flood of new mathematics, the people who can judge where it should point become economically valuable, because curating and directing a large body of new math is a much more levered move than before.
“those who understand it and are able to make the decision of where it should point actually have a lot more economic value by being able to make that judgment as a curator”
The incentive in science must shift from merely proving things to consolidating proofs into understandable problems or higher-level insights, since exposition is what makes results valuable to humans.
“Or the incentive would have to change, not just in mathematics but in other areas of science, from proving things about the world to consolidating proofs into problems or higher-level insights.”
Checking whether a math proof is correct is much easier to automate than judging good writing, because a proof is either correct or not, while writing quality involves judgments that AI verifiers get wrong, often preferring uninsightful text.
“the step to simply know if a proof is correct or not lends itself to an automated verifier, even in natural language.”
Because connections are harder to score than solved problems, AI-driven progress in connection-finding will require a lot of human-in-the-loop judgment about what connection was being sought, and most useful progress over five years will look like filling in that landscape.
“But I think it will require a lot more 'human in the loop' to say, 'What was the kind of connection that we were going for?'”
Patel admits he is moving the goalpost: finding a counterexample to the unit distance conjecture was the kind of cross-domain connection that AI should have been able to make, so the next benchmark needs to be harder.
“So it’s total goalpost moving. But then we can ask, what is the next benchmark?”
Over the next five years it would be disappointing and surprising if AI math progress produced no economically valuable improvements in areas like engineering and materials science, because applied groups can evaluate and use the insights.
“It’d be a little bit disappointing and a little bit surprising if there weren’t, over the next five years, economically valuable improvements”
A quoted maxim ranks mathematicians: good ones prove theorems, great ones come up with conjectures, and the greatest come up with definitions, implying conjecture-generation and definition-generation are the premium skills.
“good mathematicians prove theorems, great mathematicians come up with conjectures, and the greatest mathematicians come up with definitions.”
Mathematics has many episodes where results were proven long before they were understood, so the gap between reach and grasp is a recurring feature of the field and a key concern for AI-generated proofs.
“There are plenty of instances in the recent history of math where it feels like the reach has exceeded the grasp, where things are proven long before they're understood.”
The continuum hypothesis is independent of standard axioms, so whether a size of infinity exists between the naturals and reals depends on the axiom system, and forcing is the difficult method used to show this.
“The answer is both yes and no. It depends on your axioms.”
The unit distance conjecture counterexample released by an AI company included a chain of thought in natural language with no Lean, showing that natural-language reasoning can produce mathematically accepted advances.
“Why is AI able to disprove the conjecture about the unit distance problem? They released the chain of thought, or at least a rewrite of the chain of thought. It didn’t have any Lean in it.”
Mathematical discovery is mostly a random walk of being wrong, so even digesting an AI-produced proof that is known to be correct would feel like real progress because it confirms the direction leads to a solution.
“Mostly it feels like a random drunken walk, where you’re doing a thing and then you’re wrong and constantly discovering that.”
In a world where AI proves Millennium Prize problems and invents new fields, the most valued human job would be distilling what AIs have learned, because abundance and demand for explanation would make that role valuable.
“In that world, if there are any jobs whatsoever, surely distilling what the AIs have learned will be one of them.”
A seemingly hard IMO problem that stumped many strong students, including Terence Tao, has a simple brute-force-like solution that people miss because they are anchored to contest-style solution expectations, so escaping context is what unlocks it.
“There’s this almost brain-dead solution that is the best.”
Eric Jang's experience with an AI Go bot showed that autonomous agents are good at pursuing an experiment down one path but bad at stopping at dead ends and running extremely parallel exploration.
“It’s really good at running an experiment and going down that path, but it’s bad at stopping at dead ends and doing extremely parallel things.”
Diversifying AI reasoning requires a clear ontology of distinct thinking approaches, such as prove-versus-disprove and then many tactics, so that old-school software can amplify entropy by running each approach separately.
“If you’re able to put a clear ontology to the distinct ways of thinking that you want to prompt, you explore that full ontology”
The IMO's problems are designed to resist memorized training, yet a large share of them can be trained for, so solving them is weaker evidence of creativity than the problem designers intended.
“The dirty secret with the IMO is that you really can train for a lot of them.”
Human curators and musicians retain value even when machine output is objectively as good, because the social story behind a work is part of what audiences value.
“It’s the same reason human musicians are always going to have a role: that social function of the story behind them”
Institutes produce breakthroughs through serendipitous conversation among co-located experts, and AI systems could be engineered to replicate that by having agents with different expertise interact deliberately.
“We all know that an institute is smarter than an individual.”
Patel finds that for much consumed writing, pasting it into an LLM for explanation yields better results than the human-written text, so his revealed preference favors LLM distillation even though LLMs are called bad writers.
“The explanation will be better than the thing produced by the human.”
Extending Mathlib autonomously is a metaphor for civilization, since humanity's accumulated knowledge is now distilled into models that will extend it arbitrarily.
“I also love this extension of Mathlib as a metaphor for what’s going to happen to our civilization pretty soon.”
An unsupervised AI mathematical library would need a supervisor model giving heuristics about usefulness, because otherwise it would random-walk through logical space instead of producing interesting directions.
“You’d probably want some supervisor model that’s trying to provide heuristics on whether it’s useful or not.”
Interviews with mathematicians from mid-2025 to 2026 show a tone shift in how they discuss AI, which Sanderson treats as the most natural measure of AI's growing usefulness for research direction.
“it’s fun to see a little bit of a tone shift in the way they talk about AI between mid-2025 and where we are now in 2026.”
Dwarkesh admits his earlier question about AI connecting ideas was moving the goalpost, since a moderately intelligent person with that much knowledge would be expected to make such connections, and the new benchmark must be defined differently.
“So it's total goalpost moving. But then we can ask, what is the next benchmark?”
Combinatorics problems, which are playful and puzzle-like, are where AI still struggles at the IMO, and those problems are the strongest candidate for requiring genuine creativity.
“But it struggles on those combinatorics ones.”
Sanderson expects to keep doing his current work until he dies, arguing that explanation that builds lasting understanding and motivation is more durable than a one-off answer, as with the saying that a fire warms for a night but setting a man on fire warms him for life.
“You build a man a fire, and he's warm for one night. But set a man on fire, and he's warm for the rest of his life.”
Sanderson noted that the IMO was old news and not worth revisiting, since AI has already passed it, so it adds little to the current discussion.
“If we just think about IMO, which is old news at this point.”